Fuzzy Epistemicism∗

John MacFarlane†

October 9, 2008

Abstract

It is taken for granted in much of the literature on vagueness that se-

mantic and epistemic approaches to vagueness are fundamentally at odds. If

we can analyze borderline cases and the sorites paradox in terms of degrees

of truth, then we don’t need an epistemic explanation. Conversely, if an

epistemic explanation suffices, then there is no reason to depart from the fa-

miliar simplicity of classical bivalent semantics. I question this assumption,

showing that there is an intelligible motivation for adopting a many-valued

semantics even if one accepts a form of epistemicism. The resulting hybrid

view has advantages over both classical epistemicism and traditional many-

valued approaches.

∗I presented versions of this paper in June 2007 at the Arche Vagueness Conference in St. An-drews, Scotland, and the LOGICA Conference in Hejnice, Czech Republic. I am grateful toaudiences at both conferences for their comments, and particularly to Dorothy Edgington, mycommentator at St. Andrews. I would also like to thank Branden Fitelson, Michael Caie, Fab-rizio Cariani, Elijah Millgram, Stephen Schiffer, Mike Titelbaum, and two anonymous referees foruseful correspondence.

†Department of Philosophy, University of California, 314 Moses Hall, Berkeley, CA 94720-2390. Email: jgm@berkeley.edu.

1

It is taken for granted in much of the literature on vagueness that semantic and

epistemic approaches to vagueness are fundamentally at odds. If we can analyze

borderline cases and the sorites paradox in terms of degrees of truth, then we don’t

need an epistemic explanation. Conversely, if an epistemic explanation suffices,

then there is no reason to depart from the familiar simplicity of classical bivalent

semantics. Thus, while an epistemic approach to vagueness is not logically in-

compatible with the view that truth comes in degrees, it is usually assumed that

there could be no motivation for combining the two.

My aim in this paper is to question this assumption. After describing the

way in which many-valued theories are usually motivated in opposition to epis-

temicism (§1), I give an argument for degrees of truth that even an epistemicist

should be able to accept (§2). Unlike traditional motivations for degree theories,

this argument is compatible with the epistemicist’s claim that we are irremedia-

bly ignorant of the semantic boundaries drawn by vague terms, and with nonse-

mantic (epistemicist and contextualist) approaches to the sorites paradox. Thus it

opens up conceptual space for a hybrid between fuzzy and epistemic approaches,

a “fuzzy epistemicism.” According to fuzzy epistemicism, both uncertainty and

partial truth are needed to understand our attitudes towards vague propositions. In

§3, I consider how this hybrid theory can respond to some traditional objections

to many-valued theories.

I do not think that this all adds up to a compelling case for fuzzy epistemicism

as the best approach to vagueness. As I will indicate, there are a couple of non-

epistemicist approaches that seem at least equally promising. My aim here is to

2

show that if one is inclined towards epistemicism, then (contrary to the conven-

tional wisdom) one has good reason to accept degrees of truth as well.

1 The standard dialectic

If “tall man” has a classical extension,1 then there is a shortest tall man. Of course,

we have no way of knowing how tall the shortest tall man is. And even if we

could know, the placement of the line between the tall and the non-tall would

appear arbitrary. Unlike “gold” and “water”, “tall” does not seem to pick out

any kind of natural property. Nor does anything about our use of “tall” make any

particular cut-off point salient. So classical semantics is committed to unknowable

and arbitrary-seeming semantic boundaries.

Epistemicism is an attempt to bite this bullet, by explaining on general episte-

mological grounds why we should expect to be ignorant in just this way, and by

rejecting as verificationist the idea that we should be in a position to know exactly

where the semantic boundaries lie. According to the epistemic approach, what

distinguishes vague language from non-vague language has nothing to do with

truth-conditions. Formally, then, epistemicism is compatible with both classical

and nonclassical semantics. Typically, however, epistemicists defend classical se-

mantics.

One popular alternative to classical semantics is to suppose that truth comes

in degrees. The most common form of this view represents these degrees by real

1Relative to a context and an index of evaluation. I will not repeat this qualification in whatfollows.

3

numbers between 0 and 1, with 1 representing complete truth, 0 complete falsity,

and the intermediate values various degrees of “partial truth.” The extensions of

predicates are then naturally understood as fuzzy sets, or mappings from objects

to degrees of truth. Thus, “tall man” may map a 7-foot man onto 1, a 6-foot man

onto 0.75, a 5-foot-11 man onto 0.68, and so on. Small differences in height will

yield small differences in the degree to which the predicate is satisfied. So as we

look at shorter and shorter men, we will see a slow, steady decline in the degree to

which “tall man” applies, rather than a sudden, precipitous change from inclusion

in the extension to non-inclusion.

Such a theory affords an attractive analysis of the sorites paradox. Suppose

we have a line of 100 men of gradually increasing height. Man 0 satisfies “tall

man” to degree 0, man 1 to degree 0.01, man 2 to degree 0.02, and so on up to

Man 100, who satisfies “tall man” to degree 1. Now consider the following sorites

argument:

(1) Man 100 is a tall man.

(C100) If Man 100 is a tall man, Man 99 is a tall man.

(C99) If Man 99 is a tall man, Man 98 is a tall man....

(C1) If Man 1 is a tall man, Man 0 is a tall man.

(2) Therefore, Man 0 is a tall man.

4

On the Łukasiewicz semantics for the conditional, ~A→ B� = 1 if ~B� > ~A� and

1 − (~A� − ~B�) otherwise (where p~φ�q denotes the degree of φ). So all of the

conditionals C1. . .C100 have degree 0.99. That is, they are all almost completely

true, and that, the degree theorist proposes, is why we are inclined to accept them.

But although modus ponens is valid in the sense of preserving degree 1, it is

not valid in the sense of preserving degree of truth in general. Thus, when the

premises of a modus ponens inference do not all have degree 1, the conclusion

can have a lower degree than any of the premises. With each application of modus

ponens, then, we lose a little truth, so that by the end of the argument we have

none left at all.

Notice how the degree theory is motivated as an alternative to epistemicism.

By positing a smooth continuum of partial truth, we avoid the need to explain

how our linguistic practices could fix a sharp boundary between the tall and the

non-tall, and why we could not know where it lies. And by making it possible to

say that the premises of the sorites are almost completely true, we avoid the need

to explain why we should be inclined to accept a conditional that is just plain false

(as one of C1. . .C100 must be, if classical semantics is correct).

The standard epistemicist response to such theories is to argue that they merely

put off the pain, because the epistemicist’s resources will be needed anyway, at a

later stage of analysis. So if the point of degree theories is to avoid having to tell

epistemic stories, these theories are unmotivated. Let us look at some arguments

to this effect.

5

1.1 Hidden boundaries

One of the things that seemed objectionable about classical semantics was its com-

mitment to unknowable, arbitrary-seeming semantic boundaries. But do degree

theories do better? Just as on classical semantics, there will be a shortest man

who falls into the extension of “tall man,” so on a many-valued semantics, there

will be a shortest man who satisfies “tall man” to degree 1. A man 1 mm shorter

than this man will not satisfy “tall man” to degree 1. We have no way of knowing

where this boundary lies, and even if we could know it, it would seem arbitrary.

So the degree theory does not have any evident advantage over classical semantics

in this respect. Roy Sorensen puts the point effectively:

. . . advocates of alternative logics that use the sensitivity objection

against the epistemic approach are guilty of special pleading. Given

that the super-valuationists and many-valued theorists cannot use the

sensitivity issue to claim an advantage over classical logic, what is

left to recommend their positions? The central motive for appealing

to these alternative logics was to avoid the commitment to unlimited

sensitivity. Once it is conceded that this appeal cannot succeed, there

is no longer any point in departing from classical logic. (Sorensen

1988, 247)2

2See also Keefe 2000, 115: “The best epistemic theorists offer detailed explanations of why weare ignorant in a borderline case. . . ; a degree theorist taking option (i) similarly owes us an expla-nation of the ignorance it postulates, but one that does not at the same time justify the epistemictheorist’s position about first-order vagueness. It is far from clear that this can be done.”

6

Degree theorists standardly respond that their precise assignments of degrees

are meant as models of something imprecise. The sharp boundaries, they say, are

just artifacts of the numerical models being used (Edgington 1997, 297, 308–9;

Cook 2002). This is a plausible response, but more must be said. Degree theorists

ought to say which features of their models are artifacts, and which are meant to

represent real features of degrees of truth (Keefe 2000). An obvious thought is

that the ordering of the numerical degrees represents the real ordering of degrees

of truth, even if it is an artifact which degree is represented by the number 0.5. But

if the ordering is non-artifactual, so is the boundary between the maximal degree

and all the others. So, also, is the question which which of a series of successively

taller women satisfies “tall” to a greater degree than Sarah satisfies “short.”

Indeed, as Rosanna Keefe points out, the degree theorist cannot coherently

hold that only ordinal relations between numerical degrees represent relations be-

tween real degrees of truth. For the Łukasiewicz semantics for the conditional

makes the ordinal position of conditionals depend on the absolute difference of

the numerical degrees of their antecedents and consequents (Keefe 2000, ch. 5).

So if we have conditional propositions, then the absolute distances between nu-

merical degrees cannot be artifacts of the model unless some facts about ordering

are also artifacts.

A natural proposal, explored by Cook 2002, is that only large differences

in numerical degree represent real differences in degrees of truth (cf. Edgington

1997, 297–8). As Cook shows (244), this proposal is not strictly tenable: for ex-

ample, on Edgington’s theory, if there are n mutually independent propositions,

7

there will be at least some non-artifactual differences in degree less than or equal

to 1/2n, and for plausibly large values of n, these differences will be very small.

Importantly, though, these small differences will be knowable in principle, since

they can be predicted from the semantics of the connectives. So perhaps it is a suf-

ficient reply to the epistemicist’s tu quoque about unknown and arbitrary semantic

boundaries to say that

. . . truth (and falsity) do come in gradations, and both large differences

in real number assignments and the logical relations between complex

sentences and their constituents are indicative of real aspects of vague

natural language. On the other hand, the assignment of particular real

numbers to particular sentences, and the resulting sharp boundaries,

are just conveniences, incorporated into the semantics for the sake

of simplicity, but reflecting nothing actually present in the discourse

being modeled. (Cook 2002, 245)

As Cook notes, to say this is not to make the semantics itself imprecise, since the

word “large” is used not in the formal semantics, but in our informal description

of how the semantics models linguistic reality. The fit between a formal model

and the reality it models should not be expected to be precise.

I won’t try to assess this response here. What’s important for our purposes is

that both sides in the debate assume that, if the numerical degrees are viewed in

a strongly representational way, and not as models with many artifactual features,

then degree theory is unmotivated. Both sides agree that if we are going to accept

hidden and arbitrary-seeming semantic boundaries, we might as well stick with a

8

bivalent semantics. That is why the degree theorist must parry the classicist’s tu

quoque by adopting the modeling perspective.

1.2 The sorites

It might be thought that the attractive many-valued analysis of the sorites para-

dox provides an independent reason for preferring many truth values to two. But

on closer examination, this apparent advantage evaporates. As Weatherson 2005

observes, the sorites is no less compelling when run with negated conjunctions

instead of conditionals:

(1) Man 100 is a tall man.

(NC100) It’s not the case that Man 100 is a tall man and Man 99 is not.

(NC99) It’s not the case that Man 99 is a tall man and Man 98 is not....

(NC1) It’s not the case that Man 1 is a tall man and Man 0 is not.

(2) Therefore, Man 0 is a tall man.

But with the usual many-valued semantics for the connectives,3 (NC50) gets de-

gree 0.5—meaning that it is no more true than false. What this shows is that we3~P & Q� = max(~P�, ~Q�) and ~¬P� = 1 − ~P�. This is the semantics that is usually dis-

cussed in the philosophical literature on degree theories (e.g. in Machina 1976, Williamson 1994,and Keefe 2000). Different choices are made in the fuzzy logic literature (see Hajek 2006). In“Łukasiewicz logics,” strong conjunction is defined as follows: ~P & Q� = max(0, ~P�+~Q�−1).If the conjunctions in our sorites are understood this way, the NCi’s will all have degree 0.99.However, as we will see in the next section, there are strong reasons (independent of the sorites)for the degree theorist not to define conjunction this way. See note 5, below.

9

can’t hope to explain the plausibility of the sorites argument solely by pointing to

the very high degree of truth of its premises, since only in the conditional version

of the argument do all the premises have a high degree of truth.

This is not to say that a degree theorist can’t explain the plausibility of the

sorites—just that the explanation cannot advert to the “near complete truth” of the

premises. Weatherson endorses Kit Fine’s suggestion that we are prone to confuse

P with pDeterminately Pq, even when P occurs as part of a larger sentence. So

we take the (NC50) to be true because we conflate it with

(NCd50) It is not the case that Man 50 is determinately tall but Man 49 is determi-

nately not tall.

But as Weatherson notes, “Fine’s hypothesis gives us an explanation of what’s

going on in Sorites arguments that is available in principle to a wide variety of

theorists”—supervaluationists, classical semanticists, and degree theorists alike.

As a result, a degree theorist who makes use of this explanation cannot claim to

have an advantage over any of these other theories in explaining the plausibility

of sorites arguments.4

Other explanations of the pull of the sorites are also possible. Perhaps we

mistake our inability to give a counterexample to (NC50) for evidence of its truth.

Williamson 1994 argues that, because of general “margin of error” requirements

on knowledge, we could never know that we had a counterexample (234). Contex-

tualists argue that active consideration of a particular height changes the context

4Weatherson, who is arguing for a kind of degree theory himself, concedes that he doesn’t“have a distinctive story about the Sorites in terms of truer.”

10

so that the extension of “tall” draws no boundaries there (Raffman 1996; Soames

1999; Fara 2000). Either of these strategies might explain why we are unable to

refute (NC50), and hence why it seems plausible.

There is no reason why a degree theorist couldn’t appeal to these explanations

of the plausibility of the sorites. But then the degree theorist’s semantics would

not be doing any work in explaining the apparent force of sorites arguments. So,

one wonders, why not just stick with the simpler classical semantics?

To sum up: the usual motivations for a degree-theoretic account of vague ex-

pressions assume that epistemic accounts of the sorites and of borderline cases

are untenable. Both sides in the debate agree that if the degree theorist were to

accept the epistemicist’s explanation of our ignorance of the locations of sharp

semantic boundaries, the game would be lost. They agree that there would be

no point being an epistemicist and accepting a many-valued semantics, since the

epistemicism would deprive the many-valued semantics of any useful job to do.

2 A new argument for degrees

Having brought this assumption into the open, I now want to question it. I will

present a new argument for a many-valued semantics for vague discourse. Unlike

the standard motivations for degree theories, this one is compatible with epis-

temicism and does not require a “modeling” perspective on numerical degree val-

ues. The core of the argument is an acute observation by Schiffer 2003. Though

Schiffer himself rejects degree theories and argues instead for a complex “psy-

11

chological” theory, I will argue that the position that Schiffer’s observation really

supports is a degree theory that accepts hidden semantic boundaries—a hybrid of

traditional degree theories and traditional epistemic theories.

2.1 Combining uncertainties

Consider Borderline Jim. He’s just short of six feet tall, with a small tuft of hair on

his head, and he’s pretty fast at solving sudoku puzzles, though not as fast as his

brother Bill. He is, we might say, borderline tall, borderline bald, and borderline

smart. Given Jim’s borderline status, it would be wrong for us to flat-out believe

that he is tall, bald, or smart. But it would also be wrong to flat-out believe that he

is not tall, not bald, or not smart. The appropriate attitude is something between

full acceptance and full rejection, though what kind of attitude is less clear.

Classical semantics would seem to commit us to a particularly simple an-

swer to this question. Since according to classical semantics, there are facts of

the matter as to whether Jim is tall, bald, or smart, our attitude toward each of

these propositions should be one of uncertainty. If Jim is a paradigm borderline

case—right in the middle between clear satisfiers of these predicates and clear

non-satisfiers—we might take it to be 50% likely that Jim is bald, 50% likely that

he is tall, and 50% likely that he is smart. Rather than full belief, we will have

partial beliefs—credences of 0.5—in each of these propositions.

But what should our attitude be to the conjunction of these propositions? As-

suming (harmlessly, I think) that these propositions are stochastically indepen-

dent, our credence in the conjunction ought to be the product of our credences

12

in the conjuncts: 0.125. Classical semantics, then, recommends that we should

endorse conjunctions of independent borderline propositions much less strongly

than we endorse the conjuncts individually.

But, as Schiffer observes, this just seems wrong (Schiffer 2003, 204). It seems

perfectly appropriate to endorse the conjunctive proposition that Jim is tall and

bald and smart to about the same (middling) degree as we endorse the conjuncts

separately. Certainly it seems wrong that we should be quite confident (0.875)

that Jim doesn’t have all three properties.

If you don’t have these intuitions, try increasing the number of independent

properties. With seven independent properties, your credence that Jim has all of

them should be less than 0.01, and your credence that Jim doesn’t have all of them

greater than 0.99. That is, if Jim is also borderline fat, borderline old, borderline

rich, and borderline nice, you should be very confident that he is not tall, bald,

smart, fat, old, rich, and nice. Are you?

The argument, then, runs as follows:

1. If classical semantics is correct for vague discourse, then borderline propo-

sitions are either true or false; no finer distinctions are made.

2. If borderline propositions are either true or false, then (since we don’t know

which truth value they have) our attitudes toward them must be attitudes of

uncertainty-related partial belief.

3. If our attitudes towards borderline propositions are attitudes of uncertainty-

related partial belief, they ought to obey norms of probabilistic coherence.

13

4. We regard the propositions Jim is tall, Jim is bald, and Jim is smart as

independent. That is, we don’t think Jim’s being bald (or smart, or bald and

smart) would make it any more likely that he is tall, and so on.

5. Probabilistic coherence demands that our credence in the conjunction of

several propositions we take to be independent be the product of our cre-

dences in the conjuncts.

6. But it is not the case that we ought to have much less credence that Jim is

bald and tall and smart than we have that he is bald.

7. Therefore, classical semantics is not correct for vague discourse.5

Unlike the usual arguments against classical semantics for vague discourse,

this argument is not aimed at the classicist’s commitment to unknowable and

arbitrary-seeming semantic boundaries, and it has nothing to do with sorites argu-

ments. Instead, it is aimed at the idea that our attitude toward borderline proposi-

tions is one of uncertainty as to whether they are true or false.6

5A similar argument can be used to rule out many-valued theories in which conjunction isunderstood as Łukasiewicz “strong conjunction” (see note 3, above). On such theories, P & Q & Rwill have degree 0 when P, Q, and R each have degree 0.5. So this kind of fuzzy theorist will beeven less well placed than the classical logician in accounting for our partial endorsement of theconjunction.

6Sorensen seems to reject the intuition that supports premise (6). He argues as follows againstdegree theories: “. . . suppose a speaker begins by describing Ted as short and then adds that he isalso fat, bald, smart, athletic, and rich. We assign a degree of truth of 0.5 to ‘Ted is short’ and0.6 to each of the remaining attributions. But contrary to the conjunction rule [of many-valuedsemantics], we do not believe that ‘Ted is short, fat, bald, smart, athletic, and rich’ equals thedegree of truth of ‘Ted is short.’ Our uncertainties compound making us assign a much lowerdegree of truth to the claim that Ted exemplifies the conjunctive predicate. . . . Also notice that‘Ted is fat, or bald, or smart’ is less of a borderline attribution than ‘Ted is fat.’ ” (Sorensen 1988,235–6). Note that this argument just assumes that the degrees represent “uncertainties,” which the

14

One might try to defend classical semantics by rejecting (2). This is essentially

what Schiffer does. (Though he does not present his view as a way of defending

classical semantics, he emphasizes that it is a psychological solution to the sorites,

and is thus at least consistent with classical semantics.) Schiffer argues that our

attitude to borderline propositions is not standard uncertainty-related partial be-

lief (SPB), but a special kind of vagueness-related partial belief (VPB): “It is a

primitive and underived feature of the conceptual role of each concept of a vague

property that under certain conditions we form VPBs involving that concept, and

it is in this that vagueness consists” (Schiffer 2003, 212). VPBs are distinguished

from SPBs in the following ways (198–207):

• SPBs represent uncertainty, while VPBs represent ambivalence.

• SPBs generate corresponding likelihood beliefs, while VPBs do not. If one

has a SPB of 0.5 that one left one’s glasses at the office, one will take it to

be 50% likely that one’s glasses are at the office. But if one has a VPB of

0.5 that Jim is bald, one will not take it to be 50% likely that he is bald.

• Generally, if one has an intermediate SPB that p, one thinks that one is

degree theorist ought to deny.An alternative way of rejecting (6), suggested by an anonymous referee, would be to acknowl-

edge the intuitions that are taken to support it, but claim that they are misleading and not to be takenas normative. Psychologists have shown that ordinary intuitions about probabilities frequently vi-olate even the most basic norms of probabilistic coherence: in one famous case, a majority ofsubjects took a conjunction to be more likely than one of its conjuncts (Kahneman and Tversky1983; for a different interpretation of the data, cf. Crupi et al. forthcoming). Could it be that theintuitions to which Schiffer has drawn our attention are the result of the “conjunction fallacy” orsomething similar? That seems unlikely, since these intuitions can be found even in those who arenot prone to probabilistic fallacies when vagueness is not in play. But there is room for furtherempirical investigation here.

15

not in the best possible epistemic position to pronounce on p. But one can

have an intermediate VPB that p and think that one could not be in a better

epistemic position to pronounce on p.

• SPBs are governed by norms of probabilistic coherence, whereas VPBs are

governed by the Łukasiewicz many-valued truth tables. Thus, if one has a

VPB of 0.5 that Jim is bald and a VPB of 0.5 that Jim is tall, one ought to

have a VPB of 0.5 that Jim is bald and tall, even when the conjuncts are

independent.

It’s this last feature that allows Schiffer’s theory to say that our degree of belief

that Jim is tall and bald and smart shouldn’t be less than our degree of belief in

any of the conjuncts singly, when Jim is a borderline satisfier of each predicate.

Schiffer insists, reasonably, that

(*) SPB(p) + SPB(¬p) + VPB(p) + VPB(¬p) = 1.

Where p is a complete borderline case, SPB(p) and SPB(¬p) will both be 0, and

VPB(p) and VPB(¬p) will sum to 1; where p is fully determinate, the VPBs will

be 0 and the SPBs will sum to 1. But mixed cases are also possible, and on these

Schiffer’s theory runs aground. Suppose, for example, that you think there’s a

50% chance that Sam is completely hairless and a 50% chance that he has about

50 hairs on his head. (You can’t remember which of two men he is.) If you

knew he was completely hairless, you’d have an SPB of 1 that Sam is bald. If

you knew that he had 50 hairs, you’d have a VPB of 0.8 that Sam is bald, and

16

of 0.2 that he is not bald. But given your uncertainty, you’re in a mixed state,

with some SPB and some VPB in both the proposition that Sam is bald and its

negation. Schiffer gives some plausible principles for computing SPBs and VPBs

in cases like this, but as I show in MacFarlane 2006, they are inconsistent with

(*).7 The basic problem should be evident: the norms governing SPBs and VPBs

are fundamentally different, so they are not going to march in the kind of lockstep

that would be needed to keep them summing to 1.8

2.2 Taking-to-be-partially-true

Let us return to the problem Schiffer’s theory was supposed to solve. Some

kind of partial or qualified endorsement seems appropriate for borderline proposi-

tions. However, this partial endorsement does not seem to be standard uncertainty-

related partial belief, since if it were, the degree of endorsement would drop dra-

matically as we added independent conjuncts. How, then, should we understand

it?

Here, at last, we have a task degrees of truth are well suited to perform. My

proposal, to simplify slightly, is that we understand this partial endorsement not

7The fix Schiffer proposes in his reply (Schiffer 2006) does not work. In fact, the first coun-terexample in MacFarlane 2006—SPB(p) = VPB(p) = SPB(q) = VPB(q) = 0.3, SPB(¬p) =

VPB(¬p) = SPB(¬q) = VPB(¬q) = 0.2—is a counterexample to Schiffer’s revised proposal aswell, and it is easy to generate others.

8An alternative approach, due to Hartry Field (2003), is to avoid positing VPBs but allowSPB(p) + SPB(¬p) < 1. In cases we take to be completely indeterminate, SPB(p) + SPB(¬p) willbe 0. Field’s approach agrees with Schiffer’s in predicting that one should have the same degree ofbelief in the proposition that Jim is tall and tall and smart that one has in the conjuncts separately,but disagrees about what this degree should be—for Field, it is 0. Schiffer objects (210 n. 38) thatagents should not have the same degree of belief (0) in propositions they take to be borderline asthey do in propositions they take to be determinately false.

17

as partial belief in the truth of a proposition, but as belief in its partial truth. That

is not quite the right thing to say, as it makes the attitude seem like a thought about

a proposition, not about (say) Jim. In addition, it makes it seem as if the attitude

requires deployment of a concept of degrees of truth—a concept many believers

lack. But just as we might usefully understand first-order belief as taking-to-

be-true, so we might understand the first-order partial endorsement appropriate

in borderline cases as taking-to-be-partially-true (for example, taking-to-be-true-

to-degree-0.5). In describing the attitudes this way, we identify them by their

constitutive aims. Mark Sainsbury puts the point well:

Truth is what we seek in belief. It is that than which we cannot do

better. So where partial confidence is the best that is even theoreti-

cally available, we need a corresponding concept of partial truth or

degree of truth. Where vagueness is at issue, we must aim at a degree

of belief that matches the degree of truth, just as, where there is no

vagueness, we must aim to believe just what is true. (Sainsbury 1995,

44)

An attitude towards p that a cognitive system normatively “aims” to be in just in

case p is true can justly be called “taking-to-be-true,” even if the possessor of this

attitude lacks an explicit concept of truth. Similarly, an attitude towards p that a

cognitive system normatively aims to be in just in case p is true to degree N can

be justly be called “taking-to-be-true-to-degree-N,” even if the possessor of the

attitude lacks an explicit concept of partial truth.

18

Attitudes of taking-to-be-partially-true, I suggest, can do all of the work Schif-

fer aimed to do with his VPBs:

1. They can be clearly distinguished from attitudes of uncertainty. They re-

flect, rather, ambivalence: in a case where I take p to be partially true and

partially false, I am ambivalent about whether p.

2. They fail to generate likelihood beliefs. To take p to be true to degree 0.3 is

not to take it to be 30% likely that p.

3. Taking p to be partially true is consistent with taking oneself to be in the

“best possible epistemic position to pronounce on p.” Partial truth is an

objective status, not a feature of the thinker’s mental state or epistemic po-

sition.

4. Attitudes of taking-to-be-partially-true, unlike attitudes of partial belief, are

not governed by norms of probabilistic coherence. If one takes the propo-

sitions that Jim is tall, that Jim is bald, and that Jim is smart to be true

to degree 0.5, then one should take their conjunction to be true to degree

0.5 also. (On the Łukasiewicz semantics for continuum-valued logics, the

degree of a conjunction is the minimum of the degrees of its conjuncts.)

Schiffer’s VPBs look like a way of trying to get the benefits of a degree theory

without accepting the idea that truth comes in degrees. But why not go for the

original instead of this ersatz? Schiffer offers two arguments, neither of which is

compelling.9

9Here I echo some of the discussion of MacFarlane 2006.

19

His first argument is that degree theories cannot capture what Crispin Wright

calls “the absolutely basic datum that in general borderline cases come across as

hard cases” (Wright 2001, 69–70). Schiffer argues that a degree theorist

. . . is evidently constrained to hold that p is true just in case p is T

to degree 1 (or—allowing for the vagueness of ordinary language

‘true’—to a contextually relevant high degree); false just in case p

is T to degree 0 (or to a contextually relevant low degree); and neither

true nor false just in case p is T to some (contextually relevant) degree

greater than 0 and less than 1. But suppose Harry is borderline bald.

Then, since it would be definitely wrong to say that ‘Harry is bald’ is

T to degree 1 (or to some other contextually relevant high degree), the

theory entails that it would also be definitely wrong to say it is true

that Harry is bald. But if Harry is borderline bald, it would not be

definitely wrong to say that he’s bald, and thus not definitely wrong

to say it’s true that he’s bald. (Schiffer 2003, 192)

In assuming that a degree theorist is “constrained to hold” that p is true simpliciter

just in case its degree of truth exceeds some (perhaps contextually determined)

threshold, Schiffer is thinking of a degree theory as a way of systematizing all-out

truth and falsity assignments. That is one kind of degree theory. But on the more

thoroughgoing degree theory recommended here, the degrees are given a signifi-

cance directly, not indirectly through their role in systematizing “designatedness”

or all-out truth.10 According to this theory, when it is true to degree 0.5 that Harry10Compare the discussion of M vs. MD in Weatherson 2005, §1. The fact that normal talk of

20

is bald, it will be just as correct to believe that Harry is bald as it is to believe that

Harry is not bald, and it will be just as correct to believe that it is true that Harry

is bald as it is to believe that it is false that Harry is bald. This, I think, admirably

captures the “ambivalence” we feel in borderline cases. Schiffer mischaracter-

izes this ambivalence in representing it as indecision about whether to assert the

borderline proposition. It simply isn’t correct to assert p when p is a borderline

proposition, unless one is trying to effect some kind of “accommodation” (Lewis

1979) that would make it no longer count as borderline.

Schiffer’s second argument against degree theories is that they allow that cer-

tain classically valid modes of inference (for example, reductio ad absurdum) can

take one from premises that are true to degree 1 to a conclusion that is true to a

degree very close to 0. His example is

A person with $50 million is rich.

A person with only 37¢ isn’t rich.

Therefore, it’s not the case that, for any n, if a person with $n is rich,

then so is a person with $n − 1¢.

which, on the degree-theoretic analysis, has premises true to degree 1 and a con-

clusion true to a degree slightly greater than 0. This, he says, is “apt to seem

flat-out unacceptable” (193).

But why? If we agree that the premises are true and want to reject

truth and falsity does not include degree qualifiers is no obstacle for this view, since on a naturalsemantics for “true,” “It is true that Harry is bald” will have exactly the same degree of truth as“Harry is bald.” In fact, it must have the same degree of truth if the biconditional “Harry is bald iffit is true that Harry is bald” is to get degree 1 on the Łukasiewicz semantics.

21

(C) For some n, a person with $n is rich and a person with $n − 1¢ is not rich,

then we have to give up some classically valid principle of reasoning. And a

many-valued semantics gives an illuminating story about why reductio should fail

in vague contexts. If we derive a contradiction from premises S 1, S 2, S 3 using

valid (1-preserving) inference rules, then we know that at least one of them has

degree less than 1. If we also know that S 1 and S 2 have degree 1, then we can infer

that S 3 has degree less than 1. But all we can conclude about ¬S 3 is that it has

degree greater than 0. We certainly cannot conclude that it has degree 1. That’s

why reductio fails in this context. Given that something needs to be done to block

the reasoning that leads to (C), recognizing limits on the use of reductio seems

well-motivated and at least as moderate as Schiffer’s own solution, according to

which it is indeterminate whether classical inference rules—including not just

reductio but even modus ponens—are valid (Schiffer 2003, 224).

I suggest, then, that we explicate the kind of partial endorsement that is appro-

priate in borderline cases—what Schiffer calls “vagueness-related partial belief”—

as taking-to-be-partially-true.

2.3 Combining partial truth with uncertainty

As we have seen, Schiffer’s theory founders in its attempts to integrate two sepa-

rate aspects of partiality of belief: the “ambivalence” that stems from vagueness

and the uncertainty that stems from incomplete information. Can the present ap-

proach do better in integrating taking-to-be-partially-true with partially-taking-to-

22

be-true?

This is a problem that any degree theorist must face in “mixed cases,” where

the degree of truth of a vague proposition (say, Sam is bald) depends on some non-

vague matter about which there is uncertainty (say, the number of hairs on Sam’s

head). But the problem is especially acute for theorists who view all facts about

the ordering of numerical degrees to be representationally significant (not artifacts

of the model), since on their view every attitude towards a borderline proposition

will combine ambivalence and uncertainty. We will never be in a position to

know who is the shortest man who satisfies “tall man” to degree 1, and we will

have no good basis for taking the proposition that Jim is tall to be true to degree

0.653 rather than 0.649. We may be confident that he satisfies “tall” to some

intermediate degree, and perhaps we’d bet on 0.6 over 0.5, but there will remain

some uncertainty. So, to model our attitudes towards borderline propositions, we

will need to take into account both dimensions of partiality: ambivalence and

uncertainty.

The most straightforward way to do this, I think, is to represent our attitudes

to vague propositions as probability distributions over degrees of truth (strictly

speaking, over an algebra of precise propositions that ascribe degrees of truth

to the vague propositions at issue). So, for example, your attitude towards the

proposition Jim is tall might be depicted by Figure 1, where the horizontal axis

represents degrees of truth and the height of the curve over any given degree repre-

sents the probability that Jim is tall has that degree of truth. This picture combines

both dimensions of partial endorsement, taking-to-be-partially-true and partially-

23

taking-to-be-true, in a unified representation.11

Cre

dence

0.0

1.0e-2

2.0e-2

3.0e-2

4.0e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

Jim is tall

Figure 1: Pr(~Jim is tall� = x).

This approach can deal straightforwardly with the “mixed cases” that proved

troublesome for Schiffer’s theory. Suppose you aren’t sure exactly how many

hairs Tom has on his head. Your credence function is represented by Figure 2,

where the vertical axis represents probabilities and the horizontal axis the number

of hairs.

For each possible number of hairs x, there will be a probability distribution

11This graph, and those that follow, was generated by a custom Haskell program using TimDocker’s Charts library and Martin Erwig’s Probabilistic Functional Programming library (Erwigand Kollmansberger 2006). To simplify the calculations in these charts, I use a finite set of degrees,{0, 0.01, 0.02, . . . , 0.99, 1}. This allows us to do probability calculations using simple algebra.Everything said here should generalize to real-valued degrees, but more complex methods wouldbe needed to handle probability distributions over the degrees (see Zadeh 1968).

24

Cre

dence

0.0

2.0e-3

4.0e-3

6.0e-3

8.0e-3

1.0e-2

0.0 100.0 200.0 300.0 400.0 500.0Number of hairs

Probability that Tom has x hairs

Figure 2: Pr(Tom has x hairs).

over degrees of truth that represents the attitude you would have towards the vague

proposition Tom is bald if you knew that Tom had exactly x hairs. Three of these

distributions are plotted in Figure 3.

Taking into account your uncertainty about the number of hairs Tom has on

his head, what should be your attitude towards the vague proposition Tom is bald?

Since Figures 2 and 3 both represent probability distributions, the solution is a

simple application of probability theory. We construct a probability distribution

over assignments of degrees of truth to Tom is bald as follows:

Pr(~Tom is bald� = x) =∑

0≤n<500

Pr(Tom has n hairs)× Pr(~Tom is bald� = x |Tom has n hairs)

25

Cre

dence

0.0

2.0e-2

4.0e-2

6.0e-2

8.0e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

20 hairs 250 hairs 400 hairs

Figure 3: Pr(~Tom is bald� = x |Tom has n hairs), for n = 20, 250, 400.

The resulting curve, which represents your composite attitude of partial endorse-

ment toward Tom is bald, is displayed in Figure 4.

Given probability distributions over degree-assignments for atomic proposi-

tions, we can easily calculate distributions for truth-functional compounds. This

is most straightforward in cases where the conjuncts are degree-independent:

Definition: P1, . . . , Pn are mutually degree-independent iff for all sub-

sets {P j, . . . , Pk} of {P1, . . . , Pn} containing two or more elements, and

for all 0 ≥ d j, . . . , dk ≤ 1,

Pr(~P j� = d j & . . . & ~Pk� = dk) = Pr(~P j� = d j)× · · · × Pr(~Pk� = dk).

26

Cre

dence

0.0

5.0e-3

1.0e-2

1.5e-2

2.0e-2

2.5e-2

3.0e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

Tom is bald

Figure 4: Pr(~Tom is bald� = x).

If the propositions Jim is tall, Jim is bald, and Jim is smart are degree-independent

(as seems plausible), then the probability that they will have degrees d1, d2, and

d3 respectively is just

Pr(~Jim is tall� = d1) × Pr(~Jim is bald� = d2) × Pr(~Jim is smart� = d3).

And the degree of the conjunction Jim is tall & Jim is bald & Jim is smart on

this assignment of degrees to the conjuncts is just the minimum of {d1, d2, d3}. So

we can compute the probability that the conjunction has degree x by summing the

probabilities of combinations d1, . . . , dn of degrees whose minimal member = x.

27

The result is given in Figure 5.12

Cre

dence

0.0

1.0e-2

2.0e-2

3.0e-2

4.0e-2

5.0e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

((Jim is tall & Jim is bald) & Jim is smart) Jim is tall Jim is bald Jim is smart

Figure 5: Pr(~Jim is tall and bald and smart� = x).

Our complaint about classical semantics was that it predicts that one should

have far less confidence in the conjunction Jim is tall and bald and smart than in

any of the conjuncts. The view now being considered, by contrast, predicts that

one should have a little less confidence in the conjunction than in the conjuncts.

The more uncertainty there is about the degrees of the conjuncts, the larger the

drop in confidence will be (see Figure 6). This seems to me about the right result:

midway between Łukasiewicz and Williamson.13

12Cases in which the conjuncts are not independent can also be handled, though with addedcomplexity. We consider such a case below (§3.3).

13A similar view is defended by Nicholas Smith in his contribution to this volume. The maindifference is that Smith defends the view that one’s “degrees of belief” in a proposition p should

28

Cre

dence

0.0

5.0e-3

1.0e-2

1.5e-2

2.0e-2

2.5e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

((Jim is tall & Jim is bald) & Jim is smart) Jim is tall Jim is bald Jim is smart

Cre

dence

0.0

2.0e-2

4.0e-2

6.0e-2

8.0e-2

0.1

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

((Jim is tall & Jim is bald) & Jim is smart) Jim is tall Jim is bald Jim is smart

Figure 6: Conjunction with more and less uncertainty.

29

3 Traditional objections reconsidered

Traditionally, one of the most serious problems for degree theories has been the

absence of any compelling motivation. Degrees of truth do not seem to be needed,

as some thought they were, to understand the semantics of graded adjectives,

hedges, or the ordinary predicate “true” (see Lakoff 1973, Williamson 1994, Haack

1996, Kennedy 2007). Nor is it clear that degree theories provide a better diagno-

sis of the sorites paradox than is available to the classical semanticist. Finally, it

is not clear that degree theories can avoid a commitment to arbitrary-seeming and

unknowable semantic boundaries, so if that was what was objectionable about

classical semantics, degree semantics fares no better. If we need epistemicism

anyway, why should we abandon the elegant simplicity of classical semantics?

We can now answer this question. Classical semantics should be rejected for

vague discourse because it forces us to think of the partial endorsement appropri-

ate in borderline cases as a kind of uncertainty. As Schiffer observed, this con-

ception yields implausible recommendations about our attitudes towards conjunc-

tions of independent borderline propositions. A many-valued semantics provides

be identified with the expected value of the degree of truth of p—that is, with the average degreeof truth weighted by probability—while I do not attempt to arrive at a single-number degree ofbelief. While it might be useful for some purposes to measure beliefs by the expected value ofdegree of truth, it seems rash to suppose that all the interesting quantitative differences betweenpartial belief states can be boiled down to this one number. For example, compare (a) a belief statethat assigns equal credence to each degree of truth (a flat line on our graphs), (b) a belief state thatassigns near certainty to degree 0.5 (a sharp spike), and (c) a belief state in which credence clustersaround two points, degree 0.2 and degree 0.8 (two humps with a dip in the middle). In all threecases the expected value of degree of truth will be 0.5, but these belief states can be expected tohave different effects on behavior and inference. Hence I prefer to represent states of partial beliefin two dimensions, rather than attempting to integrate the uncertainty and partial-truth aspects intoone number. That said, Smith and I agree about much more than we disagree about.

30

an elegant way to represent and work with the two kinds of partiality that can

characterize our attitudes to borderline propositions: ambivalence (taking-to-be-

partially-true) and uncertainty (partially-taking-to-be-true). It can be motivated

on these grounds even if we accept hidden semantic boundaries and a diagnosis

of the sorites paradox that is compatible with classical semantics.

But not all of the worries people have had about degree theories concern moti-

vation. It has been alleged that such theories run into problems with higher-order

vagueness, and that their use of numerical degrees involves an implausible com-

mitment to the comparability of degrees of truth. In addition, many criticisms have

been raised against degree-functional semantics for the connectives, and specif-

ically against the min rule for calculating the degree of a conjunction. In this

section, I consider how a fuzzy epistemicist might respond to these objections.

3.1 Higher-order vagueness

A classic objection to degree theories is that, even if they do give a nice story

about borderline cases and the sorites paradox, all the problems come back at a

higher level. For example, it is alleged that someone could be borderline between

satisfying “tall” to degree 1 and satisfying “tall” to a degree less than 1. If so, we

could construct a new sorites on the predicate “satisfies ‘tall’ to degree 1,” and the

degree theory would have no distinctive diagnosis of this higher-order sorites.

This is clearly not a problem for fuzzy epistemicism, which does not make use

of degrees of truth to give a diagnosis of the sorites. As we saw in section 1.2, the

many-valued analysis does not work well for sorites arguments using negated con-

31

junctions, so some other story about the sorites is needed anyway. The motivation

for degrees of truth offered in the last section is consistent with a number of differ-

ent possible accounts of the sorites paradox, including epistemic and contextualist

accounts.

Indeed, I’m not convinced that the higher-order sorites poses a serious worry

even for standard degree theories. The predicate “satisfies ‘tall’ to degree 1” is

sufficiently theoretical that it’s not clear why we should accept a sorites premise

formulated with it. Perhaps that is why the objection is usually run using a sen-

tential operator D (for “definitely”), stipulated to have the following semantics:

~Dφ� = 1 if ~φ� = 1

0 otherwise.

We do have a strong inclination to accept a sorites premise for “definitely tall.”

But it’s not clear that the ordinary meaning of “definitely” matches that of D as

defined above. More plausibly, “definitely φ” means something like “φ is true

enough, by a good margin, for present purposes,” or “φ has degree 1 by a good

margin,” and on this understanding we should expect “definitely φ” to take non-

extremal degrees, since “enough” and “good margin” are vague. If that is right,

then a degree theory can say exactly the same thing about a sorites with “definitely

tall” that it says about a sorites with “tall.”

32

3.2 Comparability

Any theory that represents degrees of truth as (real or rational) numbers, and takes

the ordering of these numbers to be significant, not an artifact of the model, is

committed to the degrees being totally ordered: for any sentences A and B, the

degree to which A is true will be either less than, equal to, or greater than the de-

gree to which B is true. Both critics and friends of degree theories have suggested

that “multidimensional predicates” pose a problem for the idea that degrees are

totally ordered.14 Here is Williamson’s version of the complaint:

Comparisons often have several dimensions. To take a schematic ex-

ample, suppose that intelligence has both spatial and verbal factors.

If x has more spatial intelligence than y but y has more verbal intel-

ligence than x, then ‘x is intelligent’ may be held to be truer than ‘y

is intelligent’ in one respect but less true in another. Moreover, this

might be held to be a feature of the degrees to which x and y are in-

telligent: each is in some respect greater than the other. How can

two real numbers be each in some respect greater than the other? On

this view, degrees are needed that preserve the independence of dif-

ferent dimensions, rather than lumping them together by an arbitrary

assignment of weights. (Williamson 1994, 131)

Multidimensional predicates do pose a problem for degree theorists who under-

stand degrees of truth in terms of comparatives. But the present theory explicates14In addition to Williamson, quoted below, see Goguen 1969, 350–1, Forbes 1985, 175, Keefe

2000, 129, and Weatherson 2005.

33

and motivates degrees of truth in terms of their role in explaining our attitude of

declining partial endorsement of successive members of a sorites series, not by

appeal to comparatives. A is F-er than B can be true to degree 1 even when A

is F and B is F have the same degree of truth (Williamson 1994, 126). Multidi-

mensional predicates just give us one more reason not to tie our understanding of

degrees of truth to comparatives.

The same considerations provide a response to this argument, from Keefe

2000:

. . . consider the case where p = ‘a is tall’ and q = ‘b is red’. Here we

have no single comparative on which ‘true to a greater degree’ can

piggy-back. The comparison may be read as ‘a is more clearly tall

than b is red’ and if, for example, a is clearly tall and b is clearly not

red, then this will be true. But in a wide variety of cases (e.g. with

a 5-foot 10-inch man and a reddish-orange patch), neither disjunct of

(CT ) [“either p ≥T q or q ≥T p”] will be true. (129)

If our sole grip on the notion of degrees of truth came from comparatives, then in

cases like these, where there are no comparatives to appeal to, there would be a

case for saying that there is no fact of the matter about whether Joe is tall is truer

than Patch #50 is red. But we have rejected the close tie between comparatives and

degrees of truth. So why think that the difficulty in determining which proposition

is truer is anything other than epistemic? Indeed, an epistemic difficulty is to be

expected on the present theory, since our attitudes towards these propositions are

34

represented as probability distributions over degrees. When their curves overlap,

we will not be in a position to know which proposition has the higher degree.

Of course, I am not opposed to the development of non-numerical degree theo-

ries that relax the requirement that degrees be totally ordered. I am just expressing

skepticism about the usual motivations for such theories. There are, in addition,

technical reasons for wanting degrees to be totally ordered. One is that it is unclear

how to define negation with partially ordered degrees (see Williamson 1994, 133).

Weatherson 2005 ends up with Boolean negation, but this is patently unsuited to

the purposes of a degree theory, as it allows the degree of A to be greater than

the degree of ¬A only when A is determinately true and ¬A determinately false.15

Surely a degree theory ought to allow that, say, Man 60 is tall can be truer than it

is false—truer than Man 60 is not tall—even if it is not completely true.

3.3 Degree functionality

Another standard group of objections to degree theories is directed at the degree-

functional semantics for the connectives. These objections are often regarded

as devastating, even by those sympathetic with degrees (most prominently, Edg-

ington 1997, who argues for a degree theory with non-degree-functional connec-

tives). Although the motivation for fuzzy epistemicism does not assume that the

connectives are degree-functional, it does assume that the degree of a conjunc-

tion of independent propositions whose degrees are 0.5 is 0.5, and none of the

15Proof: Suppose A ≥T ¬A. Then A &¬A ≥T ¬A, because ¬A ≥T ¬A, and the degree of aconjunction is the greatest lower bound of the degrees of its conjuncts under the ≥T ordering. Butthen, since A &¬A is determinately false, so is ¬A. See Weatherson 2005, 67.

35

non-degree-functional degree theories I know of deliver that result.16 But I am

not convinced that the arguments against degree-functionality are compelling, es-

pecially when one takes account of the fact that our attitudes toward borderline

propositions will generally be mixed with uncertainty.

I will focus here on negation, conjunction, and disjunction, ignoring the con-

ditional. This is fair, I think, for a couple of reasons. First, there are lots of worries

about classical logic’s truth-functional conditional, so if there are problems with

the Łukasiewicz conditional, it is not clear that they should reflect badly on the

multivalued framework. Whatever improved technologies are devised to handle

conditionals in a two-valued framework can presumably be ported to the multival-

ued framework as well. Second, the argument I’ve used to motivate degrees does

not rely on any particular semantics for the conditional. (It may be contrasted, in

this respect, with motivations for degree theory that focus on its treatment of the

sorites paradox.)

Perhaps the most obvious objection to degree functionality is that it forces us

to accept that some contradictions are not completely false. Assuming degrees of

truth make sense, it should be possible for there to be a sentence that is just as true

as it is false. Let P be such a sentence:

(1) ~P� = ~¬P�.

Now, plausibly,

(2) ~P & P� = ~P�.16On Edgington’s theory, the degree of a conjunction of degree-independent propositions is the

product of the degrees of the conjuncts.

36

It follows immediately from these premises that

(3) If & is degree-functional, then ~P &¬P� = ~P� = ~¬P� , 0.

It is hard to see how one could have a meaningful degree theory while rejecting

every instance of (1),17 and rejecting (2) would be very strange.18 Our choice,

then, is clear. We can have a degree-functional semantics for conjunction, at the

price of allowing some contradictions to be partly true, or we can ensure that all

contradictions have degree 0, at the price of rejecting degree functionality.

So can we grasp the nettle of half-true contradictions? Note, first, that a con-

tradiction that is half-true is also half-false. In accepting such things, then, one

is not committing oneself to the assertibility of any contradictions, since it is pre-

sumably not appropriate to assert half-falsehoods. One is merely accepting that

some kind of “ambivalent” attitude is appropriate towards certain contradictions.

And this doesn’t seem wrong. Indeed, “It is, and it isn’t” is just how we express

our ambivalence in borderline cases.19

Moreover, if Jim is a borderline tall man, it seems wrong to assert, “Either

he’s a tall man or he isn’t.” This is explained nicely on the hypothesis that such

disjunctions are only half-true. Granted, there are other possible explanations: for

example, we may be construing “Either he’s tall or he isn’t” as “Either he’s defi-

nitely tall or he’s definitely not tall,” or the assertion may carry some implicature17Though Weatherson 2005 seems to do just that. See note 15, above.18Though Goguen 1969, 347 does reject it, defining the degree of a conjunction as the product

of the degrees of the conjuncts. (2) is also rejected in Łukasiewicz logics with strong conjunction,where ~P & P� = ~P� only when ~P� = 1 (see note 3, above).

19When we say this, I take it, we are not asserting a contradiction; we’re finding words toexpress our attitude of partial endorsement. Note also that such uses cannot be construed as “It isin one sense, and it isn’t in another sense.”

37

about definiteness. But it should be registered that there is at least a prima fa-

cie case for assigning intermediate degrees of truth to some instances of excluded

middle (the flip-side of assigning them to some contradictions).

Finally, although many philosophers seem to think it’s obvious that a contra-

diction cannot have any positive degree of truth,20 it’s not clear what the argument

is supposed to be. Timothy Williamson writes:

By what has just been argued, the conjunction ‘He is awake and

he is asleep’ also has that intermediate degree of truth. But how

can that be? Waking and sleeping by definition exclude each other.

‘He is awake and he is asleep’ has no chance at all of being true.

(Williamson 1994, 136)

But to say that a contradiction has degree 0.5 is not to say that it has a chance of

being true. Indeed, if we are certain that it has degree 0.5, then we will take it to

have no chance of being completely true. To say that waking and sleeping exclude

one another is to say that if “x is awake” is true to degree 1, “x is asleep” is true

to degree 0, and vice versa. Perhaps it is also to say that “x is awake” is as true as

“x is asleep” is false. But on neither understanding does it rule out the possibility

that both are true to some intermediate degree.

So much for the objections to half-true conjunctions. It might be thought, how-

ever, that there are other, more powerful grounds for objecting to degree function-

ality. Here is an argument that can be found (in different versions) in Williamson

20For example, Williamson asks, “how can an explicit contradiction be true to any degree otherthan 0?” (1994, 136).

38

1994, Edgington 1997, and Keefe 2000. Let Jim be a borderline satisfier of “tall,”

so that Jim is tall is true to degree 0.5, and let Tim be just a bit shorter than Jim, so

that Tim is tall is true to degree 0.45. Then, since Jim is tall has the same degree as

Jim is not tall, by degree functionality we can conclude that Tim is tall and Jim is

not tall has the same degree as Tim is tall and Jim is tall. This is counterintuitive,

since Jim is taller than Tim. So we should reject degree functionality.

But why does it seem counterintuitive that Tim is tall and Jim is not tall and

Tim is tall and Jim is tall have the same degree of truth? It might be argued that

when Jim is taller than Tim, Tim is tall and Jim is not tall must be completely false

(degree 0). But then we must presumably say that Tim is not tall or Jim is tall is

completely true (degree 1), and that seems odd in a case where, although nothing

is hidden from us, we would not assert either that Tim is not tall or that Jim is

tall. The issues here are, I think, very similar to the issues we considered above

in connection with half-true contradictions. If we are willing to accept half-true

contradictions, it does not seem too bad to accept that Tim is tall and Jim is not

tall might be just slightly more false than true.

Even if we do accept this, though, the feeling persists that we should have

different attitudes toward Tim is tall and Jim is not tall and Tim is tall and Jim

is tall, and that the former is bad in a way that the latter is not. That is why

this objection is deeper than the bare rejection of half-true contradictions. To

resist it, we need to explain why different attitudes are appropriate towards these

propositions, despite the fact that they have the same degree of truth.

I think that fuzzy epistemicism can provide a kind of answer to this question.

39

Our attitudes towards the propositions in question will not amount to certainty

that they have such-and-such degree of truth. Rather, they will be representable

as distributions of probability over a range of possible degrees. And, as we saw

in section 2, above, attitudes toward logically compound propositions will be de-

termined by these distributions in a way that factors in both the uncertainty and

the ambivalence that they reflect. So let’s see what happens when we add a bit of

uncertainty about the degrees (using a normal probability distribution centered on

0.5 for Jim is tall and 0.45 for Tim is tall).21

Cre

dence

0.0

2.0e-2

4.0e-2

6.0e-2

8.0e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

Jim is tall Tim is tall Tim is tall & Jim is tall Tim is tall & ~(Jim is tall)

Figure 7: Differences between Tim is tall and Jim is tall and Tim is tall and Jim isnot tall when Tim is slightly shorter than Jim.

As Figure 7 reveals, there is a clear difference in the recommended distribution21Note that Jim is tall and Tim is tall are not degree-independent. It is assumed here that ~Tim

is tall� = ~Jim is tall� - 0.05.

40

of credences over degrees for our two propositions. Particularly salient is the fact

that Tim is tall and Jim is not tall cannot be truer than 0.5, while Tim is tall and

Jim is tall has a tail that goes up beyond 0.8. This, I suggest, may be enough to

account for the persistent feeling that our attitudes toward the two propositions

should differ, and that the latter is more strongly endorsable than the former.

3.4 Edgington’s argument against the min rule

The same resources can help defend against an argument, put to me by Dorothy

Edgington, that the degree of the conjunction of two independent propositions

each with degree 0.5 should be less than 0.5. The argument goes as follows.22

Suppose the degree of Jim is tall is 0.5. As we vary the degree of Jim is bald from

0 to 1, we should expect the degree of Jim is tall and bald to vary gradually from

0 to a maximum of 0.5. So the conjunction should have a degree less than 0.5

when Jim is bald has degree 0.5, contrary to the min rule.

Edgington’s assumption that the conjunction is truer when Jim is bald has de-

gree 0.7 than when it has degree 0.5 should, I think, be resisted. It draws intuitive

support from the fact that, if asked to point to the “tall, bald man,” I will point to

the balder of two equally tall men, even if both are borderline tall and borderline

bald. But this fact can be explained pragmatically; we need not conclude that

the balder man satisfies the description “tall, bald man” to a greater degree than

the thinner one. If Yao Ming and Shaquille O’Neal are both on the court, every-

22Here I draw on Edgington’s comments on an earlier version of this paper at the Arche Vague-ness Conference. For related arguments, see Edgington 1997, 304.

41

one will understand who we mean if we talk about “the tall man,” even though

presumably both of them satisfy “tall man” to degree 1.

We can do more justice to the intuitions underlying Edgington’s argument if

we bring in the dimension of uncertainty and represent the attitudes at issue by

probability distributions over degrees. As Figure 8 shows, fuzzy epistemicism

recommends a clear difference in attitude towards Jim is tall and bald as the de-

gree of Jim is bald goes from 0.5 to 0.7. The effect, as before, is due to the

uncertainties, and will be more prominent the greater the uncertainty about the

precise degree of truth of Jim is bald. Thus we can vindicate the intuitions Edg-

ington deploys against the min rule for conjunction without giving up the min rule

itself—another nice illustration of the way in which the insights of epistemicism

and degree theory can complement each other, once we give up trying to motivate

degree theory as a way to avoid hidden semantic boundaries or solve the sorites

paradox.

42

Cre

dence

0.0

1.0e-2

2.0e-2

3.0e-2

4.0e-2

5.0e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

Jim is tall Jim is bald(a) (Jim is tall & Jim is bald(a))

Cre

dence

0.0

1.0e-2

2.0e-2

3.0e-2

4.0e-2

5.0e-2

0.0 0.2 0.4 0.6 0.8 1.0Degree of truth

Jim is tall Jim is bald(b) (Jim is tall & Jim is bald(b))

Figure 8: The degree of Jim is tall and bald as the (expected) degree of Jim isbald varies from 0.5 (top) to 0.7 (bottom).

43

References

Roy Cook. Vagueness and mathematical precision. Mind, 111:225–47, 2002.

Vincenzo Crupi, Branden Fitelson, and Katya Tentori. Probability, confirmation,

and the conjunction fallacy. Thinking and Reasoning, forthcoming.

Dorothy Edgington. Vagueness by degrees. In Rosanna Keefe and Peter Smith,

editors, Vagueness: A Reader, pages 294–316. MIT Press, Cambridge, 1997.

Martin Erwig and Steve Kollmansberger. Probabilistic functional programming

in Haskell. Journal of Functional Programming, 16:21–34, 2006.

Delia Graff Fara. Shifting sands: An interest-relative theory of vagueness. Philo-

sophical Topics, 28:45–81, 2000.

Hartry Field. The semantic paradoxes and the paradoxes of vagueness. In JC Beall

and Michael Glanzberg, editors, Liars and Heaps, pages 262–311. Oxford Uni-

versity Press, Oxford, 2003.

Graham Forbes. The Metaphysics of Modality. Oxford University Press, Oxford,

1985.

J. A. Goguen. The logic of inexact concepts. Synthese, 19:325–73, 1969.

Susan Haack. Deviant Logic, Fuzzy Logic: Beyond the Formalism. University of

Chicago Press, Chicago, second edition, 1996.

44

Petr Hajek. Fuzzy logic. In Edward N. Zalta, editor, Stan-

ford Encyclopedia of Philosophy. Fall 2006 edition, 2006. URL

http://plato.stanford.edu/archives/fall2006/entries/logic-fuzzy/.

Daniel Kahneman and Amos Tversky. Extensional versus intuitive reasoning: the

conjunction fallacy in probability judgment. Psychological Review, 90:293–

315, 1983.

Rosanna Keefe. Theories of Vagueness. Cambridge University Press, Cambridge,

2000.

Christopher Kennedy. Vagueness and grammar: the semantics of relative and

absolute gradable adjectives. Linguistics and Philosophy, 30:1–45, 2007.

George Lakoff. Hedges: A study in meaning criteria and the logic of fuzzy con-

cepts. Journal of Philosophical Logic, 2:458–508, 1973.

David Lewis. Scorekeeping in a language game. Journal of Philosophical Logic,

8:339–359, 1979.

John MacFarlane. The things we (kinda sorta) believe. Philosophy and Phe-

nomenological Research, 73:218–224, 2006.

K. F. Machina. Truth, belief and vagueness. Journal of Philosophical Logic, 5:

47–78, 1976.

Diana Raffman. Vagueness and context relativity. Philosophical Studies, 81:175–

192, 1996.

45

R. M. Sainsbury. Paradoxes. Cambridge University Press, Cambridge, second

edition, 1995.

Stephen Schiffer. The Things We Mean. Oxford University Press, Oxford, 2003.

Stephen Schiffer. Replies. Philosophy and Phenomenological Research, 73:233–

243, 2006.

Scott Soames. Understanding Truth. Oxford University Press, Oxford, 1999.

Roy A. Sorensen. Blindspots. Oxford University Press, Oxford, 1988.

Brian Weatherson. True, truer, truest. Philosophical Studies, 123:47–70, 2005.

Timothy Williamson. Vagueness. Routledge, London and New York, 1994.

Crispin Wright. On being in a quandary: Relativism, vagueness, logical revision-

ism. Mind, 110:45–98, 2001.

L. A. Zadeh. Probability measures of fuzzy events. Journal of Mathematical

Analysis and Applications, 23:421–427, 1968.

46